This paper introduces SimulMask, a new paradigm for fine-tuning large language models (LLMs) for simultaneous translation. SimulMask utilizes a novel attention masking approach that models simultaneous translation during fine-tuning by masking attention for a desired decision policy. Applied to a Falcon LLM on the IWSLT 2017 dataset, SimulMask achieves improved translation quality compared to state-of-the-art prompting optimization strategies across five language pairs while reducing computational cost. Why it matters: The proposed method offers a more efficient way to adapt LLMs for real-time translation, potentially enhancing multilingual communication tools and services.
Researchers from MBZUAI have released MobiLlama, a fully transparent open-source 0.5 billion parameter Small Language Model (SLM). MobiLlama is designed for resource-constrained devices, emphasizing enhanced performance with reduced resource demands. The full training data pipeline, code, model weights, and checkpoints are available on Github.
Researchers at National Taiwan University are developing low-complexity neural network technologies using quantization to reduce model size while maintaining accuracy. Their work includes binary-weighted CNNs and transformers, along with a neural architecture search scheme (TPC-NAS) applied to image recognition, object detection, and NLP tasks. They have also built a PE-based CNN/transformer hardware accelerator in Xilinx FPGA SoC with a PyTorch-based software framework. Why it matters: This research provides practical methods for deploying efficient deep learning models on resource-constrained hardware, potentially enabling broader adoption of AI in embedded systems and edge devices.
The paper introduces Sparse-Quantized Representation (SpQR), a new compression format and quantization technique for large language models (LLMs). SpQR identifies outlier weights and stores them in higher precision while compressing the remaining weights to 3-4 bits. The method achieves less than 1% accuracy loss in perplexity for LLaMA and Falcon LLMs and enables a 33B parameter LLM to run on a single 24GB consumer GPU. Why it matters: This enables near-lossless compression of LLMs, making powerful models accessible on resource-constrained devices and accelerating inference without significant accuracy degradation.
MBZUAI researchers demonstrated a low-latency, multilingual multimodal AI system at GITEX that integrates speech, text, and visual capabilities for more lifelike human-machine conversation. The demo, led by Dr. Hisham Cholakkal, includes a mobile app where users can point their camera at an object and ask questions, receiving spoken answers in multiple languages. They are also integrating the model into a robot dog that can respond to voice commands. Why it matters: This work addresses key challenges in deploying LLMs to real-world applications in the Middle East, such as multilingual support and real-time responsiveness.